A new approach developed at MIT allows robots to think about their next movements while they are still carrying out the current ones. This leads to smoother motions and much quicker reactions. The method works by enabling the artificial intelligence (AI) model that controls the robot to forecast the machine’s future position. The model then uses that forecast to shift seamlessly from one set of actions into the next.
Many current systems force a robot to pause while it calculates what to do next. These pauses create slow and jerky movements. By basing its plans on the predicted future state of the robot rather than its present position, the MIT method removes those interruptions and lets the robot work faster. The technique adds no extra computing demand and can be used with different kinds of robotic hardware.
How the prediction works
In advanced robotics the AI systems known as vision-language-action models serve as the robot’s planning centre. A vision-language-action model takes pictures from the robot’s camera together with instructions about the task, then produces a short sequence of movements. Because the real-time process of analysing images and deciding actions is demanding, the robot often waits between sequences. The new system, called VLASH, solves this by estimating where the robot will be after it finishes its current sequence. It uses that estimate to prepare the following movements while the robot is still moving. The researchers also group movements into larger steps, a process known as action quantization, which further increases speed even though accuracy falls slightly.
In tests the method roughly doubled the speed of robots performing pick-and-place work and cut the delay between successive movements. It also improved performance in rapid activities such as table tennis and Whack-a-Mole. The approach could prove useful for robots that must move quickly in difficult settings, for example during emergency response or search-and-rescue work, and it may help robots recover more rapidly from mistakes. The research is described in a preprint and will be presented at the Intelligent Robots and Systems Conference.